About Me
A Robotics and AI student at UCL, I'm drawn to how machines perceive the world — through vision, sound, and spatial understanding.
Ningbo Xiaoshi High School
Xiamen University Malaysia · BEng Computer Science and Technology
University College London · MSc Robotics and Artificial Intelligence
Projects
SAV-SLAM: Semantic Audio-Visual SLAM
Binds sound sources to physical objects in a 3D map on commodity hardware (iPhone 16 Pro + ~$60 four-mic array). A visual branch (VGGT-SLAM + YOLO) supplies objects and geometry; a text-prompted audio branch discovers and separates emitters; a directional (DoA) layer arbitrates when the semantic prior is ambiguous, and recovers sources the visual vocabulary missed. The phone-only pipeline is demonstrated end-to-end on real recordings; array integration is under way.
VPR Loop Closure
ML / RoboticsA four-stage loop-closure front end for SLAM, built as a controlled classic-versus-modern comparison: every swap — ResNet18 to DINOv2, triplet loss to InfoNCE, ORB to DISK + LightGlue — is motivated by a measured failure of the stage it replaces. Zero-training DINOv2 features beat the fine-tuned ResNet18 under an open-set protocol (best F1 0.874 vs 0.796), and sequence re-ranking tripled Recall@100%-Precision from 0.23 to 0.69.
Footprints
Skills & Interests
Languages
- Python
- C++
- Java
Frameworks & Tools
- PyTorch
- React + FastAPI
- Docker
Robotics & Perception
- SLAM
- Computer Vision
- Multimodal Perception
- Motion Planning
- Model Predictive Control
- Kalman Filter
- Legged Robotics